Automatic Segmentation of the Abdominal Aorta and Stent-Grafts Bertram Sabrowsky-Hirsch, Stefan Thumfart, Wolfgang Fenz, and Richard Hofer Unit for Medical Informatics, RISC Software GmbH, Hagenberg, Austria Email: {bertram.sabrowsky-hirsch, stefan.thumfart, wolfgang.fenz, richard.hofer}@risc-software.at Pierre Schmit and Franz Fellner Central Radiology Institute, Kepler University Hospital, Linz, Austria Email: {pierre.schmit, franz.fellner}@kepleruniklinikum.at Abstract—Understanding and monitoring changes of the treated vessel after Endovascular aneurysm repair is crucial for the prediction of complications and risk assessment to facilitate timely intervention. Due to the complexity of the stent-graft wire frame enveloping the aortic blood lumen and the inherent artifacts caused by the metal wire, segmenting the structures required for simulation and further analysis is a non-trivial task. In this paper we present a fully automatic segmentation architecture combining two 3D U-Nets in a novel patching approach leveraging knowledge of the target anatomy. We evaluated our approach on a real world clinical dataset against a competitive baseline, yielding results that surpass the baseline in both accuracy and computation time. On our data we achieve a median Dice similarity coefficient of 0.97 for the blood lumen and 0.88 for the stent-graft segmentation. We point out two common flaws in current segmentation models: undersampling and indiscriminate patching. By addressing them appropriately, our approach gains an advantage that may benefit a multitude of segmentation tasks. Index Terms—segmentation, patch-based, centerline, U-net, stent graft, abdominal aneurysm I. INTRODUCTION Endovascular Aneurysm Repair (EVAR) was chosen for 65% of surgical interventions of Abdominal Aortic Aneurysms (AAA) between 2010 and 2013 [1], and has therefore found its place as a minimally-invasive alternative to open surgery for suited patients. EVAR greatly reduces the intraoperative stress on patients and shortens the period of convalescence. However, the procedure also entails a high reintervention rate of 20% [2], rendering postoperative monitoring indispensable. In an endeavour to improve postoperative risk assessment by predicting complications, we plan to automatically analyse blood-flow simulations based on segmentations of the treated abdominal aorta and stent-graft prosthesis (i.e., the blood lumen and wire frame). The main obstacle in streamlining and deploying such an analysis to clinical practice is the dependence on said segmentations. Manuscript received January 18, 2021; revised July 23, 2021. Computed Tomography Angiography (CTA) is acquired within the standard clinical monitoring procedure of AAA patients [3]. A practically viable workflow must therefore rely on this imaging modality for the segmentations. Creating these segmentations semi-automatically is, however, a time-consuming task due to the complex structure of the stent-graft wire frame and the imaging artifacts it introduces. Segmenting the target structures in one scan takes a trained expert between 25 and 40 minutes. In this paper, we present a method to automatically create combined segmentations of both the blood lumen and the stent-graft wire frame. A. Related Work There are several publications on the segmentation of the abdominal aorta blood lumen and stent-graft wire frame, respectively, and some of them describe fully automated methods. We are, however, not aware of any approach that encloses both segmentation tasks. As generalized approaches for blood lumen segmentation struggle due to the unique challenges introduced by the aneurysm thrombus and stent-graft wire frame, specialized methods for segmentation of AAAs and aortic dissections are better suited for the first segmentation task. While purely intensity-based methods fail due to indistinct boundaries and strong imaging artifacts, these methods often rely on graph-based techniques or deformable models. Graph-based techniques [4]-[7] utilize shape constraints to prevent leakage into neighbouring structures. The methods rely on a rough blood lumen segmentation (or centerline information [5]) that is acquired in a semi-automatic manner (e.g., using a graph- cut technique [6]) and subsequently refined. Approaches based on deformable models [8], [9] try to automatically fit contours to the target structures, but depend on seed points for the determination of the initial contour. While Kovács et al. [8] describe an automatic calculation of these seed points, their method suffers from a general lack in accuracy, especially for postoperative scans. More exotic approaches make use of radial models [10], level- set methods [11] and tracking [12], again depending on manual selection of seed points for initialization. Of the above methods, [6], [8] and [9] are the only ones tested Journal of Image and Graphics, Vol. 9, No. 3, September 2021 ©2021 Journal of Image and Graphics 67 doi: 10.18178/joig.9.3.67-73